Modeling Temporal Dynamics in Irregular ICU Data Using MWTA-LSTM

CISSOKO Mamadou Ben Hamidou, Nicolas Lachiche

Abstract:

Accurately modeling a patient’s illness and care processes is crucial in personalized predictive medicine due to their long-term temporal dependencies. However, electronic health records often contain episodic and irregularly timed data, posing a challenge for constructing personalized predictive models.

To address this, we introduce Multi-Way Adaptive Time-Aware LSTM called MWTA-LSTM, a novel deep dynamic memory neural network designed to leverage medical records and estimate illness trajectories, current states, and future risks with precision. MWTA-LSTM extends the LSTM model by parameterizing the cell state to handle irregular timing effectively and comprehend the impact of interventions on illness courses, utilizing elapsed times between events. This enhances its ability to capture temporal dynamics and accommodate variations in event and observation timing. Empirical experiments on real-world clinical datasets validate the effectiveness of MWTA-LSTM, demonstrating its superiority over existing models and robust baselines. These findings underscore MWTA-LSTM’s potential to advance personalized predictive medicine by offering a more accurate and comprehensive approach to modeling patient health trajectories.